Responses API and Codex - NexusAPI

In NexusAPI, /v1/responses is for the custom Provider configured in the Codex CLI guide, or for a model whose supported_endpoint_types in /v1/models includes openai-response.

When to use ResponsesUse /v1/responses only when supported_endpoint_types includes openai-response. If you receive status_code=500, not implemented, use the protocol listed for that model; this normally is not a temporary error to retry.

Endpoint and scope

POST https://nexusapi.link/v1/responses

Scenario	Recommended approach
Codex CLI	Use the `wire_api = "responses"` Provider configuration in the Codex CLI guide
Another app with a model whose `supported_endpoint_types` includes `openai-response`	Use that app’s Responses client or schema and verify a minimum request first
General OpenAI-compatible text app	Prefer Chat Completions
A model without `openai-response`	Use the Chat Completions or Messages protocol listed in `supported_endpoint_types`, not this endpoint

Minimum verification request

Only after model support has been confirmed, test connectivity with:

curl 'https://nexusapi.link/v1/responses' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -d '{
    "model": "YOUR_RESPONSES_MODEL_ID",
    "input": "Reply only: connection succeeded"
  }'

input is not the same field as Chat Completions messages, and the response must not be parsed as choices[0].message. Use your client’s official schema for complete fields, streaming events, and result handling.

Codex CLI configuration essentials

model_provider = "nexusapi"
model = "YOUR_MODEL_ID"

[model_providers.nexusapi]
base_url = "https://nexusapi.link/v1"
env_key = "NEXUSAPI_API_KEY"
wire_api = "responses"

Keep the key in the NEXUSAPI_API_KEY environment variable, never in config.toml or auth.json. See the Codex CLI guide for installation, OS-specific environment variables, and verification.

What to check first

The key group contains the model and the model ID exactly matches the Model Marketplace.
The client actually requests /v1/responses, not /v1/v1/....
The model’s supported_endpoint_types includes openai-response; if not, use the listed matching protocol.
Keep the Request ID, client version, and complete error text, then follow the self-check guide.